Shopify Sidekick Review: Is It Reliable for Store Decisions?

Sidekick is reliable enough for first drafts, summaries and spotting things to look at, but not for decisions on its numbers alone. Check every figure against the matching report, ask for the period and metric in the question, and review each proposed change before you publish. Treat it as a quick analyst who needs supervising.

That is a judgement, not a measurement. Nobody has published an error rate for Sidekick that I could find, and I have not run a controlled test. What exists is Shopify's own documentation of what it does and its limits, plus merchant reports of wrong answers in community threads and third-party reviews. I could not verify those reports one by one, so I will not repeat a particular claim as fact. What I can give you is a routine that tests Sidekick on your own store, in about an hour, so you rely on evidence rather than anyone's opinion, mine included.

Follow me on Instagram@sagnikteaches

What Shopify itself says about limits

Shopify's help pages are candid about boundaries, and they are a better guide to reliability than most reviews. The points that matter for decision-making:

Connect on LinkedInSagnik Bhattacharya

• Sidekick presents changes for your review before applying them. Store changes are drafts until you review each change type and click Publish all changes. A Check this before you publish banner describes extra effects.
• Published changes cannot be undone automatically. To reverse one, you ask Sidekick to change the values back, then review and publish.
• A set of store changes holds up to 50. Ask for more and it drafts the first 50 and asks you to publish or discard them before the rest. Drafts expire 10 days after the last update.
• It works within your staff permissions, so it cannot show you Shopify Payments information you cannot already see.
• It cannot release reserves, remove account holds, change payout details or explain failed payouts.
• It cannot change your privacy policy, or any policy with Use automated policy turned on. For other policies it writes into the policy editor, and nothing changes until you click Save, so the wording is yours to check.
• It does not decide whether your store must register for a sales-type tax or owes it. For that, Shopify says to contact your tax authority or a tax professional.
• For data questions it can write and run code to forecast, compare or summarise store data. On some stores it cannot run code.
• Before using a third-party app it asks your approval, once or with "don't ask again", and with memory off it asks every time.
• Store changes cannot be reviewed or published in the Shopify mobile app; Shopify says to use Sidekick on desktop or in a mobile browser for that.

Read that list as a map of where the designers think the risk is: money movement, policies, tax and bulk changes. It also shows where they think the risk is low enough to automate: drafting, reviewing, explaining. Your own routine should follow the same split.

Subscribe on YouTube@sagnikteaches

Three kinds of answer, three levels of checking

Not every Sidekick output deserves the same suspicion. Sort what you ask into three kinds.

The middle row is where wrong answers do the most damage, because a number looks authoritative. A plausible but mistaken figure about last month's sales can drive a stock order or a discount. Two sibling tutorials cover the general version of this problem: AI hallucinations explained for business owners and how to catch made-up figures in AI drafts.

A one-hour test on your own store

The most useful thing you can do is test Sidekick where you already know the answer. Pick ten questions whose answers you can find in your reports in under a minute each, and ask them in plain words. Keep the log in a spreadsheet: question, Sidekick's answer, the report's answer, match or not, notes.

Make the set cover different kinds of question:

1. A simple total: "What were my total sales last month?"
2. A ranking: "Which five products sold the most units in the last 90 days?"
3. A ratio: "What was my conversion rate last week?"
4. A comparison: "How did this month's orders compare with the same month last year?"
5. A customer question: "How many customers bought more than once in the last 12 months?"
6. A stock question: "Which products have fewer than 10 units on hand?"
7. A question about a quiet period: "What happened on the day I ran no ads?"
8. A question with a trap, such as a period that straddles a month end.
9. A question about something that does not exist: "How did my wholesale channel perform?" when you have none.
10. A question about Shopify itself: "Does Shopify charge a fee for X?", which you check against the help page.

Question nine tells you how it behaves when it should say "I have no data". A tool that invents a wholesale channel is a tool you must watch more closely. Question ten tests general knowledge, which is not tied to your store data and where its answer should match Shopify's own documentation.

Main example: a cleaning company with a small online shop

Consider an illustrative cleaning company that sells refill concentrates and cloths through a Shopify store, alongside its contract work. About 140 orders a month. The owner runs the ten-question test on a Tuesday evening. Here is what the log looks like for four of the ten (all figures illustrative).

Three lessons come out of that table, and they are typical of the kind of thing to expect.

First, "sales" is a loose word. The report shows total sales and net sales, and they differ because of returns, discounts and shipping treatment. If Sidekick's figure matches one and your decision needs the other, you have a problem that looks like a match. Fix: put the measure in the question ("net sales after returns and discounts").

Second, the conversion rate mismatch is not necessarily an error. Different tools count sessions differently, and a rate over a different window will differ. The note to write is not "wrong" but "ask which report and window it used, then reproduce that figure". If you cannot reproduce it, do not use it.

Third, the invented wholesale trend is the serious one. The prompt contained a false premise and the answer accepted it. That is the behaviour to design around: never ask a leading question whose answer would feel good, and do not accept analysis of anything you cannot see in the reports yourself.

How to ask so the answer is checkable

A good question names the metric, the period and the comparison, and asks Sidekick to show its working. Compare:

WEAK
How are sales doing?

STRONGER
Using net sales (after returns and discounts), give me total net
sales for 1 to 31 October, and for 1 to 30 September. Show the two
figures and the percentage difference. Say which report or
ShopifyQL query you used, and list any assumption you made about
refunds or time zones.

The stronger prompt does three jobs. It removes ambiguity, so the figure can be compared with a report. It forces the assistant to say how it got there, which gives you something to check. And it surfaces assumptions that would otherwise be hidden, such as time zone. Shopify's help page says Sidekick can generate custom ShopifyQL queries in the ShopifyQL query editor and display report visualisations, so asking it to show the query lets you open and run that query yourself.

A realistic reply to the stronger prompt (illustrative): "October net sales: $6,985. September net sales: $6,210. Difference: +12.5%. Based on the Sales report, net sales, store time zone. Assumed refunds are counted in the month they are issued." You then open the report, filter to the same range, and compare. A mismatch of a few dollars points to an assumption (refund timing, say), and a mismatch of hundreds points to a wrong measure or period.

Checking changes: what to look at before you publish

Shopify's design helps you here, because Sidekick drafts changes and waits. Your part is to read them properly. A quick routine for any change set:

1. Before approving, write down the current values you care about: prices, tags, titles, policy text. A screenshot is enough.
2. Ask for a small set first, such as five products, not the full catalogue. The 50-change limit is a cap, not a target.
3. Read every draft line. For price changes, check the arithmetic on two or three items by hand.
4. Read the "Check this before you publish" banner. It describes effects beyond the obvious ones.
5. Publish, then open three affected pages on your storefront as a customer would.
6. Wait a day and look at the orders for anything odd, such as a price that looks wrong.

A short illustrative example of why: the owner asks Sidekick to add the tag "refill" to all products with "concentrate" in the title. It drafts 23 tag changes. Reading the list, she notices two products where the word appears in an unrelated context, a "concentrate-free" cloth. She removes them from the draft before publishing. That took thirty seconds and avoided a collection page showing the wrong item. Without the read-through, the tag would have gone live on products where it is wrong.

The forecast that carried a one-off order forward

An owner asks Sidekick to forecast next quarter's sales for the refill range. It writes and runs code over past orders, and returns a smooth curve with a figure at the end. He orders stock on that basis. What he did not notice is that the data had a spike from a one-off bulk order by a single customer, so the forecast carried that spike forward as a trend. The assistant did not flag it, because he never asked.

The fix is to ask for the forecast with its assumptions and to look at the raw series first. Prompt: "Show me monthly units for the last 12 months as a table, flag any month more than 30% above the average, then forecast the next three months with and without those months." If the two forecasts differ a lot, the decision is sensitive to the outlier, and you should talk to the bulk buyer before ordering. A forecast is a story about the past told forward, and it needs a human to ask whether the past was typical.

Three cross-checks for Sidekick's quieter errors

The round-trip check

Ask the question a second way and see whether the answers agree. If Sidekick says repeat customers made 31% of orders last quarter, ask for the count of orders from customers with more than one order in the period, and the total orders, and divide them yourself. An answer that survives two phrasings is more likely to rest on the data. One that changes with the wording is a sign it is guessing at a definition.

The reasonableness check

Before opening a report, say out loud roughly what you expect: about 140 orders at about $50 should be near $7,000. If Sidekick's answer is $70,000 or $700, you do not need a report to know something is off. Owners who know their business have a built-in sense of scale, and it catches decimal slips and wrong-period answers faster than a spreadsheet.

The permissions check

Because Sidekick works within staff permissions, two people can get different answers to the same question if their access differs. If a team member reports a figure that does not match yours, compare who asked before you suspect the tool. Keep one account for the weekly numbers and note it in your log.

What to do when it gets something wrong

Treat a wrong answer as information. First, record it in your log with the exact wording of your question. Second, rephrase with the metric, period and source stated, and see whether the corrected version matches. Third, if you were about to act on the wrong figure, trace where else you used it: a stock order, a note to a supplier, a figure in a report to a partner. Fourth, rate the answer: Sidekick shows thumbs-up and thumbs-down icons on each response, and you can add a comment. Shopify says the feedback helps improve the quality of responses, which is not the same as fixing your answer today, so treat it as a ten-second courtesy rather than a correction.

For changes that went live and should not have, the route is the one Shopify describes: ask Sidekick to put the values back, review, and publish. Your pre-change note of the old values is what makes that a five-minute job and not an archaeology project. For orders, remember that edits apply only after you click Update order, so cancelling out before that point leaves the order as it was.

When to rely on it, and when to hold back

For a wider comparison of when to use Sidekick and when to reach for a general assistant, see best AI assistant for Shopify by job. Once a prompt has passed your checks several times, you can save it, and Sidekick skills for repeat tasks shows how, so the checked wording is the one you reuse.

Keeping a running scorecard

One test in November does not tell you about March. Models and features change, so repeat a shorter version each quarter. Keep a sheet with three columns per question: date, matched or not, cause of any mismatch. After a few rounds you see patterns specific to your store, such as "ratios are right but rankings differ on ties" or "it forgets the time zone". Those patterns are more useful than any general verdict, including mine, because they describe the behaviour of the tool on your data.

Add a second scorecard for changes: how many drafts you accepted unchanged, how many you edited, how many you discarded, and whether you reversed any after publishing. If you accept every draft without a single edit for months, either the tool is excellent for your tasks or nobody is reading the drafts, and a spot check of five published changes will tell you which. Nobody publishes a benchmark for a healthy edit rate, so compare yourself with your own earlier quarters rather than a figure from elsewhere.

If you want help designing the checks for the decisions that matter most in your business, that is something my AI implementation consultation covers, and you can reach me here.

The hour to spend this week

Write ten questions you can answer from your reports, ask them in Sidekick in plain words, record the answers beside the real figures, and mark each match, mismatch or invention. Then rewrite the mismatched ones with the metric and period spelled out and try again. You will then know, for your own store, which kinds of question to trust, which to check and which to avoid. That is a better basis for deciding how much to rely on Sidekick than any review, and it costs one evening.

Questions about trusting Sidekick

Does Sidekick make things up?

Like any AI assistant it can state wrong things confidently. Merchants in community discussions and third-party reviews report errors, though I could not verify individual claims. The sensible assumption is that any figure or factual claim needs checking against your reports or Shopify's help pages before you act on it.

Will Sidekick publish changes without asking?

Shopify's help page says store changes are drafts until you review each change type and click Publish all changes. Form changes are highlighted for review before you save, and order edits apply only after you click Update order. Read each draft before you publish.

Can I undo a Sidekick change?

Not automatically. Shopify says published changes cannot be undone with a single action; you ask Sidekick to change the values back, then review and publish again. Some changes cannot be fully reversed, for example when the item changed again after you published. Note the original values before approving anything that affects prices, policies or navigation.

Should I trust Sidekick on tax questions?

No. Shopify's own help page says Sidekick does not decide whether your store must register for a sales-type tax or owes it, and tells you to contact your tax authority or a tax professional. Treat any tax answer as a prompt for a question to your adviser.

Is a wrong answer more likely with certain questions?

Vague or multi-part questions, and anything that turns on dates, definitions or calculations, leave an AI assistant more room to guess than a narrow, well-defined request. Shopify's own help page says naming the exact metrics and columns you want gives a more useful result than a general question. Specify the period, the metric and the measure, then compare the answer with the matching report.

Further reads

• How to Use Shopify Sidekick to Run Your Store Faster — Everyday store jobs Sidekick handles well.
• Shopify Sidekick Pulse: Which Recommendations to Act On — Score the recommendation cards on your Home page.
• AI for Ecommerce Analytics: Shopify Reports, GA4 or ChatGPT First? — Where to look first when you check a number.
• Stocky Alternative for Shopify: Purchase Orders and AI Reordering — Reorder stock with purchase orders in the Shopify admin.
• Online Shop AI Mistakes That Hurt Trust and Conversions — AI mistakes that cost shoppers' trust.
• Does Shopify Have AI? What Is Built In and What Costs Extra — Which Shopify AI features come with Basic, Grow, Advanced and Plus, at what price, and where the extra costs begin.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Shopify's Sidekick help pages on getting help and guidance (store changes, Shopify Payments, policies, orders, analytics, third-party apps), setting up and using Sidekick (limitations, feedback) and the Sidekick overview page. All read 5 November 2026.